After AI Finds a Candidate, What Gets Paid For?

Anthropic has announced that Claude identified a distinctive combination of a reverse transcriptase, repeat sequences and other elements in existing sequence data. The function of this system, named ART, remains unknown. This is not a completed gene-editing technology, but another candidate worth investigating experimentally. Official announcement ↗

The next step involves more than isolating the sequence. What reaction does it perform? What does it act on, and what does it leave untouched? Does it offer an advantage over existing tools? Only after answering these questions can we begin to consider its uses.

As a business, this work can take two distinct forms.

One is to perform experiments for a fee. Emerald Cloud Lab, for example, provides experimental capacity with charges for instrument time, materials, storage and other resources. Synthesizing, expressing and measuring candidates proposed by AI can fit into this model. Official pricing documentation ↗

The other is to turn a candidate into something ready for further development, while retaining rights to the resulting asset.

We explored this distinction in “What Scientific AI Companies Sell Next: Experimental Data, Labs, and Drug Discovery Assets” ↗, comparing Ginkgo’s experimental services with Schrödinger’s drug discovery business.

Schrödinger does more than license software. Its agreements with Novartis combine expanded software access with joint drug discovery. The discovery collaboration provides for an upfront payment, potential future milestone payments and royalties. Agreement announcement ↗

When Tectora was established this September, Schrödinger contributed two early-stage drug discovery programs in exchange for an equity stake and potential future payments. The new company raises development funding, so Schrödinger does not bear every cost itself. Alongside fees for using its computational tools, it retains a stake in the value of the programs those tools help produce. Tectora announcement ↗

How much evidence does a candidate need before it attracts the next investment?

In the neuroscience collaboration between Recursion and Genentech, predicted targets underwent successive tests of pathways, cellular function and disease-related phenotypes. Genentech then exercised a contractual option to move a target into joint early-stage drug discovery. This is not yet clinical evidence of a drug’s efficacy, but it was enough for the partner to commit to the next stage of development. Collaboration update ↗

The story is more complicated than more AI-generated candidates creating more demand for experiments. Better predictions could also reduce unnecessary experiments. Nor does processing more samples necessarily translate into high profits for an experimental service provider.

The more important distinction is between completing a requested measurement and assembling the evidence needed for the next development decision. A company that takes responsibility for the latter and secures rights to the resulting asset can earn returns that are not determined solely by the volume of work performed. In exchange, it also takes on the risk that development will fail.

ART’s next step lies here too. Can a report of an unusual sequence become a demonstration of a function that someone wants to use? Who will undertake that work, and what will they receive in return? It is at this stage that the differences between scientific AI businesses become clear.